MeshAnything provides an official implementation for generating 3D meshes from various inputs using autoregressive transformers, targeting researchers and developers in 3D computer vision and graphics. It enables the creation of artist-quality meshes, enhancing existing 3D reconstruction pipelines.

How It Works

The project leverages autoregressive transformers to generate 3D meshes, tokenizing mesh components and predicting them sequentially. This approach allows for detailed and structured mesh generation, similar to how artists create models. It builds upon concepts from MeshGPT and Michelangelo, incorporating vector quantization for efficient representation.

Quick Start & Requirements

• Installation: pip install git+https://github.com/buaacyw/MeshAnything.git or clone and install dependencies (torch==2.1.1, torchvision==0.16.1, torchaudio==2.1.1 with CUDA 11.8, flash-attn).
• Prerequisites: Ubuntu 22, CUDA 11.8, A100/A800/A6000 GPU recommended. Python 3.10.13.
• Demo: Local Gradio demo available via python app.py.
• Usage: Command-line inference for mesh or point cloud inputs (python main.py ...).
• Examples: Hugging Face Spaces

Highlighted Details

• Generates meshes with up to 800 faces.
• Inference takes ~30s on an A6000 GPU with ~7GB VRAM.
• Supports mesh and point cloud (with normals) inputs.
• Recommends using high-quality inputs from 3D reconstruction or SDS methods.

Maintenance & Community

• The project has a newer version, MeshAnything V2, released on June 17th, which supports up to 1600 faces.
• Based on several popular open-source repositories.

Licensing & Compatibility

• The repository does not explicitly state a license. Based on its dependencies and common practice for research code, it is likely intended for research purposes. Commercial use may require clarification.

Limitations & Caveats

The current version is limited to generating meshes with fewer than 800 faces and requires sharp input geometry for optimal results. The project's README does not specify a license, which could impact commercial adoption.